IN-PLANE MATERIAL CONTINUITY FOR THE DISCRETE MATERIAL OPTIMIZATION METHOD

IN-PLANE MATERIAL CONTINUITY FOR THE DISCRETE MATERIAL OPTIMIZATION METHOD René Sørensen1 and Erik Lund2 1,2 Department of Mechanical and Manufacturing Engineering, Aalborg University Fibigerstraede 16, DK-9220 Aalborg East, Denmark 1 Email: rso@m-tech.aau.dk, web page: http://www.m-tech.aau.dk Keywords: Discrete material optimization, In-plane material filter, Laminate composite structures ABSTRACT When performing discrete material optimization of laminated composite structures, the variation of the in-plane material continuity is typically governed by the size of the finite element discretization. For a fine mesh, this can lead to designs that cannot be manufactured due to the complexity of the material distribution. In order to overcome this problem, engineers typically group elements together into socalled patches which share design variables. However, because the shape and size of a patch are fixed during the optimization procedure, a poor patch layout may drastically limit the design space, resulting in suboptimal designs. In this work, in-plane material filters are applied for controlling the material continuity. Here, the engineers can specify a minimum length scale that governs the smallest variation in the material. With this approach, the optimizer is free to determine which material to apply together with its shape and location. 1 INTRODUCTION Laminated composites are today applied in a wide variety of structures ranging from badminton rackets and bicycles to airplanes and wind turbine blades. When designing laminated composite structures, engineers have to determine which material candidate to apply, its orientation, and number of layers, throughout the entire structure. Because of the complex relationship between the structural response and the applied material layup, determining the right configuration can become an iterative and time-consuming task. In recent years, researchers have developed several methods that can aid engineers in determining a suitable material layup. Because manufacturers typically rely on a discrete set of material candidates, methods for performing discrete material optimization has seen much interest in the scientific community, see e.g., [1] and [2] for a comprehensive review on optimization of laminated composite structures. In the Discrete Material Optimization (DMO) method by [3], and [4] the material candidates are interpolated using continuous design variables. This approach thus allows for the application of efficient gradient-based optimizers. Later, [5] extended the DMO method by introducing multi-material variations of the well-known SIMP and RAMP interpolation schemes applied for traditional topology optimization, see [6] and [7], respectively. Because the analysis is typically conducted using the finite element method, the parameterization

R. Sørensen and E. Lund or distribution of the design variables to the analysis model is usually done on element level i.e., each finite element is assigned a series of candidate material variables which the optimizer can manipulate. However, when doing so the optimized layup may contain variations in the material or fiber orientations that are too complex to manufacture. In order to avoid these complicated designs, engineers typically group elements together into so-called patches that share the same design variables. The engineers thus effectively enforce a fixed minimum length-scale upon the optimized design. However, because the patch layout is fixed the optimizer cannot change neither the shape nor location of the patch. Consequently, great insight to the problem at hand is required in order to obtain designs with good performance. In this work, in-plane material filters are applied for controlling the material continuity. Here, the engineers can specify a minimum length scale that governs the smallest size a material candidate can attain. With this approach, the optimizer is free to determine which material to apply together with its shape, size, and location. 2 METHOD 2.1 Parameterization In the context of laminated composite structures, the parameterization or distribution of the design variables to the analysis model is typically done on either element or patch level i.e., groups of elements. Ideally, for a given layer in a given shell finite element the material properties can be determined using the so-called candidate material variables which in binary terms are defined as (1) For a total of material candidates, each candidate may represent a discrete set of material properties such as discrete orientations of an orthotropic material, or a core material such a foam or balsa wood when designing sandwich structures. However, in order to apply efficient gradient based optimizers, the integer requirement is relaxed and the design variables thus treated as continuous variables. In this work, the multi-material interpolation method by [5] is applied to interpolate between the material candidates. For stiffness optimization, the constitutive properties for each layer in each element can be interpolated as (2) In order to force the optimizer to select just one material, non-discrete valued design variables are penalized using the RAMP scheme. Here, for the weight factor associated with each material

candidate,, is reduced if, and thereby favoring binary valued design variables. The resource constraint requiring the design variables to sum to unity ensures that an increase in one variable must be followed by a reduction in the others. Consequently, the optimizer cannot select all material candidates at the same time, and thereby create superior but non-physical pseudo-material. 2.2 In-plane material filters With the above formulation the design variables are distributed on element level which can lead to optimized designs with small and complex variations in the material distribution. In traditional topology optimization with isotropic materials, a minimum length-scale for the material continuity can be obtained by the application of a linear density filter, see [8], and [9] (1) Here, is the volume of the l'th layer for the i'th element within the filter radius, r, and linear decaying weight function defined as is a (2) where, is the center coordinates of the i'th element. For curved structures, the distance between elements should be determined such that it follows the contours of the shell. However, for plates the Euclidean norm between element coordinates is correct. The linear density filter has previously been applied by [10] in the context of discrete material optimization of composite beam cross sections properties. By applying a linear density filter on the material candidate variables, the optimized design will always contain a relative large amount of non-discrete design variables. In order to comprehend this, the filtered design variables are passed through a threshold projection filter by [11] which is a continuous approximation of the Heaviside function. (3) Here, defines the projection limit, and controls the steepness of the Heaviside approximation. In Figure 1, the threshold projection filter is shown for (, ) = (5,0.75), (10,0.85), and (20,0.95).

R. Sørensen and E. Lund Figure 1: Threshold projection filter shown (, ) = (5,0.75), (10,0.85), and (20,0.95) Because the projection filter is a non-linear function in the design variables, the sum of design variables may not equal unity after the projection i.e.,.the optimizer thus has the opportunity to create superior but non-physical pseudo-materials. In order to avoid this problem, the projected design variables are passed through a final so-called normalization filter which is defined as (4) With the above definitions of the candidate material variables, the constitutive properties for a single layer in a given shell element can be interpolated as follows (5)

2.3 Penalization and continuation In order to favor discrete design variables, non-discrete variables are penalized using the RAMP scheme. The penalization is gradually increased using a so-called continuation strategy where the penalization is increased each time the optimization procedure converges to an optimum. In this work, the penalization is increased three times using the following sequence (1) Likewise, the and parameters applied in the projection filter are increased each time the optimization procedure converges. The parameters are increased using the following sequences (2) 3 NUMERICAL EXAMPLES The numerical examples consist of a single layered, square, corner hinged plate subjected to a uniform pressure of 1Pa. The plate is modeled using nine node equivalent single layered shell elements. The analysis is solved using linear assumptions. The length and width of the plate are 1m and the total thickness is 1mm. A sketch of the plate with boundary conditions is shown in Figure 1. The finite element model consists of 100 100 shell elements in an ordered mesh. z y x Figure 1: Sketch of plate with boundary conditions The applied material candidates all consist of Glass Fiber Reinforced Polymer (GFRP) plies with four discrete orientations {-45º,0º,45º,90º}. The objective is to minimize the total compliance of the structure. The material data is shown in the table below. E11 E22 E33 G12 G23 G13 ν12 34.0GPa 8.20GPa 8.20GPa 4.50GPa 4.00GPa 4.50GPa 0.29 Table 1: Constitutive properties for the applied Glass Fiber Reienforced Polymer

R. Sørensen and E. Lund Results from three examples will be presented in the following. For Example 1, no in-plane material filters are applied, so to obtain a reference design without any restrictions with respect to the material continuity. For Example 2, the in-plane material filters are enabled with a radius of m. For Example 3, the elements have been patched together in a 5x5 grid where each patch covers 20x20 elements. This configuration thus gives an equivalent minimum length scale as specified for Example 2 i.e., 0.2m. Because the design variables are continuous, it is possible that some variables may be non-discrete at the optimum solution. In order to quantify the amount of non-discreetness in the design, the measure of candidate non-discreteness by [12] is applied and repeated here for convenience. (10) The measure of candidate non-discreteness is normalized such that it yields 100% if all design variables have a value of, and 0% if discrete 0/1 variables have been obtained. In the case of nondiscrete designs, rounded results are also presented. The rounding of the design variables is conducted as follows (11) 4 RESULTS 4.1 General comments The results from the three examples are summarized in the table below. Compliance [J] Design iterations Example 1: No filters 0.854406E-02 0.000 16 Example 2: r=0.2m Original results 1.024250E-02 2.118 31 Example 2: r=0.2m Rounded results 0.848692E-02 0.000 - Example 3: 5x5 patches 0.857632E-02 0.000 11 Table 2: Results for the three examples As can be seen from the Table 2, Example 1 converged in just 16 design iterations and all design variables obtained discrete values by means of penalization alone. For Example 2, the measure of nondiscreteness was just above 2%, showing that the presented method is indeed capable of obtaining near discrete designs despite the application of the filters. Notice that the rounded design obtained a lower compliance value when compared to Example 1. Consequently, the results from Example 1 must represent a poor local minimum. This is indeed a possible outcome due to the non-convex design

space associated with the applied penalization scheme. The poorest performance was obtained for the patch design, however, the design variables did converge to discrete values by means of penalization alone. 4.2 Example 1 Figure 2 shows the optimum fiber angle distribution for Example 1 i.e., no filters applied. From the edges towards the center of the plate, the material layout is quite regular, however, near the center the layout becomes more complex due to the variations in the fiber orientations. Still, notice that the design is almost symmetric except near the center of the plate. 4.3 Example 2 Figure 3 shows the optimum fiber angle distribution for Example 2 i.e., r=0.2m. Compared to the previous example, the same tendencies in the fiber orientations are observed. However, the design is considered to be more manufacturable. The width of the plies reaching from the four corners towards the center of the plate is now almost constant and equal to the applied filter radius. Likewise, at the center of the plate the material now consists of pure plies. 4.4 Example 3 Figure 4 shows the optimum fiber angle distribution for Example 3 i.e., the patch design. Compared to Example 2, the optimized fiber distribution shows the same tendencies, however, because of the patch approach, the change in fiber orientation is much more abrupt. Figure 2: Example 1, no filters applied. Optimum fiber distribution Figure 3: Example 2, filter radius r=0.2m. Optimum fiber distribution Depicted circle have an diameter of d=0.2m

R. Sørensen and E. Lund Figure 4: Example 3, Patch design. Optimum fiber distribution 5 CONCLUSIONS In this work, in-plane material filters have been introduced for the discrete material optimization method. The in-plane material filters make it possible for engineers to specify a minimum length scale for the material continuity in the optimized design. Traditionally, engineers had to group elements together into so-called patches in order to control the material continuity. However, as patches are fixed the optimizer cannot influence the size, shape nor location of the patches. Compared to an equivalent patch model, the presented results for the in-plane material filters obtained a better performance and had a simpler material layout. REFERENCES [1] H. Ghiasi, D. Pasini, and L. Lessard, Optimum stacking sequence design of composite materials Part I - Constant stiffness design, Compos. Struct., vol. 90, no. 1, pp. 1 11, 2009. [2] H. Ghiasi, K. Fayazbakhsh, D. Pasini, and L. Lessard, Optimum stacking sequence design of composite materials Part II: Variable stiffness design, Compos. Struct., vol. 93, no. 1, pp. 1 13, 2010. [3] J. Stegmann and E. Lund, Discrete material optimization of general composite shell structures, Int. J. Numer. Methods Eng., vol. 62, no. 14, pp. 2009 2027, 2005.

[4] E. Lund and J. Stegmann, On structural optimization of composite shell structures using a discrete constitutive parametrization, Wind Energy, vol. 8, no. 1, pp. 109 124, 2005. [5] C. F. Hvejsel and E. Lund, Material interpolation schemes for unified topology and multi-material optimization, Struct. Multidiscip. Optim., vol. 43, no. 6, pp. 811 825, 2011. [6] M. P. Bendsøe, Optimal shape design as a material distribution problem, Struct. Optim., vol. 1, no. 4, pp. 193 202, 1989. [7] M. Stolpe and K. Svanberg, An alternative interpolation scheme for minimum compliance topology optimization, Struct. Multidiscip. Optim., vol. 22, no. 2, pp. 116 124, 2001. [8] T. Bruns and D. a Tortorelli, Topology optimization of non-linear elastic structures and compliant mechanisms, Comput. Meth. Appl. Mech. Eng., vol. 190, no. 26 27, pp. 3443 3459, 2001. [9] B. Bourdin, Filters in topology optimization, Int. J. Numer. Methods Eng., vol. 50, no. 9, pp. 2143 2158, 2001. [10] J. P. Blasques and M. Stolpe, Multi-material topology optimization of laminated composite beam cross sections, Compos. Struct., vol. 94, no. 11, pp. 3278 3289, 2012. [11] F. Wang, B. S. Lazarov, and O. Sigmund, On projection methods, convergence and robust formulations in topology optimization, Struct. Multidiscip. Optim., vol. 43, no. 6, pp. 767 784, 2011. [12] S. N. Sørensen and E. Lund, Topology and thickness optimization of laminated composites including manufacturing constraints, Struct. Multidiscip. Optim., vol. 48, no. 2, pp. 249 265, 2013.